Development of a Composite Pain Measure for Persons with Advanced Dementia: Exploratory Analyses in Self-Reporting Nursing Home Residents

Abstract

Context

Experts agree that pain assessment in non-communicative persons requires data from sources that do not rely on self-report, including proxy reports, health history, and observation of pain behaviors. However, there is little empirical evidence to guide clinicians in weighting or combining these sources to best approximate the person’s experience.

Objectives

The aim of this exploratory study was to identify a combination of observer-dependent pain indicators that would be significantly more predictive of self-reported pain intensity than any single indicator. Because self-reported pain is usually viewed as the criterion measure for pain, self-reported usual and worst pain were the dependent variables.

Methods

The sample consisted of 326 residents (mean age: 83.2 years; 69% female) living in one of 24 nursing homes. Independent variables did not rely on self-report: surrogate reports from certified nursing assistants (CNA IPT), Checklist of Nonverbal Pain Indicators (CNPI), Cornell Scale for Depression in Dementia (CSDD), Pittsburgh Agitation Scale (PAS), number of painful diagnoses, and Minimum Data Set (MDS) pain variables.

Results

In univariate analyses, the CNA IPT scores were correlated most highly with self-reported pain. The final multivariate model for self-reported usual pain included CNA IPT, CSDD, PAS and education; this model accounted for only 14% of the variance. The more extensive of the two final models for worst pain included MDS pain frequency, CSDD, CNA IPT, CNPI and age (R² = 0.14).

Conclusion

Additional research is needed to develop a predictive pain model for nonverbal persons.

Introduction

Self-report is widely accepted as the gold standard for measuring pain in research and clinical care (1–4). For persons with cognitive impairment secondary to dementia, however, this standard often is unattainable. An individual’s decreased ability to provide self-report contributes to the well-documented deficiencies in the assessment and management of pain for this vulnerable group (5–9).

Alternate methods to assess pain in nonverbal persons with dementia have been developed and tested. One method is observation of behaviors associated with pain, such as moaning, combativeness, facial grimacing, and guarding painful body parts. Several observational tools have been developed; however, psychometric testing for each of them is limited and there is insufficient evidence for experts to recommend one tool for all clinical settings and situations (10–14). Moreover, some tools lack sensitivity because they focus on too few behaviors, and others lack specificity by including behaviors associated with general discomfort and agitation and not exclusively with pain (11). The tools also cannot reliably discriminate between mild, moderate and severe pain (10, 11).

Obtaining proxy reports from health care providers and family caregivers is another approach to measuring pain. Although this approach is used in clinical settings and research, many studies demonstrate a lack of concordance between proxy reports and self-report of pain (1, 15–19).

Clues about the presence of pain come also from reports demonstrating an association between pain and other clinical states. For example, well-established associations exist between pain intensity and higher levels of depressive symptoms in older persons (20–25). Agitation also is positively correlated with pain in persons with cognitive impairment (26–28). Agitation and depression, however, may exist independent of pain and be associated with other conditions. For example, agitation in a person with advanced dementia may reflect over-stimulation or other environmental disturbance (29, 30). Medical conditions such as arthritis, cancer, pressure ulcers and recent fracture also are significantly associated with pain presence and intensity in older adults (31, 32), although there often is poor concordance between diagnoses and pain. For example, many people with radiographic evidence of osteoarthritis do not experience significant pain or disability (33).

Inadequacies in the current pain measurement methods for persons with advanced dementia require novel approaches to overcome them. Several authors have recommended using a combination of pain-related indicators to better assess and measure pain (1, 2, 10, 34). Despite these proposals, there is little empirical evidence to guide pain assessment using multiple indicators in persons with advanced dementia.

The purpose of this study was to conduct exploratory analyses to develop a composite pain measure for persons with advanced dementia using observer-recorded scales. Because self-report commonly is the gold standard for measuring pain, these initial analyses tested combinations of pain-related variables to predict self-reported usual and worst pain intensity in a sample of verbal nursing home residents with pain.. Our goal was to explore whether or not a combination of pain indicators would be significantly better in predicting self-reported pain intensity than any single pain indicator.

Methods

Participant Recruitment

Participants were recruited as part of an ongoing randomized controlled trial comparing two methods of training and supporting nursing home staff about pain assessment and management. This study included data from 24 facilities that were recruited into the parent study. All facilities were located in Western Washington State; nine were not-for-profit, 14 were for-profit, and one was a government (Veterans) facility. Inclusion criteria were: age 65 years and older, identified as having moderate to severe pain, and expected to be at the facility for at least six months.

Residents with pain were identified in one of three ways. First, research staff interviewed the unit manager or resident care manager (licensed nurses who oversee the care of residents on a unit) and asked them to identify all residents on the unit roster they believed had moderate to severe pain at any time in the past week that was not adequately treated by current therapies. Second, research staff obtained from the Minimum Data Set (MDS) Coordinator or medical records department the names of all residents who had a “2 = moderate pain” or “3 = times when the pain is horrible or excruciating” marked on Section “J2b” of the MDS (35). Third, medical charts for all other residents were reviewed to identify those with clinical notes about pain, analgesic use, or pain care plans. Residents with medical records reflecting evidence of pain were further screened for eligibility.

All study procedures were approved by the Swedish Medical Center Institutional Review Board, in Seattle, WA. Residents provided written consent or were consented by the designated healthcare proxy noted in the medical record. If the research or nursing staff were unsure whether or not the resident was capable of self-consent, or if the resident or decision-maker requested dual consent, both the resident and proxy provided written consent.

Conceptual Framework and Study Measures

The analytic model for the study (Figure 1) was an adaptation of two frameworks: “Hierarchy of Assessment Techniques” (HAT) and Snow et al.’s conceptual model (34). Originally described by McCaffery and Pasero (36) and expanded by Herr et al. (1), the HAT calls for using five categories of assessment approaches for persons with limited ability to communicate: self-report (which should be elicited whenever possible), identification of pain-related conditions and procedures, evaluation of behavioral pain indicators, surrogate reports of pain, and response to empirical analgesic trials. For the current study, we used self-report as the criterion measure and did not include a measure for response to empirical analgesic therapy. We also drew on the work of Snow et al. (34), who proposed testing a multidimensional model of pain for persons with dementia by: 1) using multiple measures that are appropriate both for cognitively intact and impaired persons; 2) adapting existing self-report measures for use by informants who provide proxy reports; and 3) adapting existing measures used for persons with dementia to capture information about behaviors, conditions and outcomes that are associated with pain, e.g., affect and agitation, which the authors refer to as “known correlates” of pain, Initial validation of the model involves an approach that the authors call the “self-report validation” paradigm in which a pain variable, based on external ratings of pain, is compared to pain self-reports from persons who are cognitively intact (34).

Figure 1.

Conceptual Model

Data sources included chart review, resident interviews, certified nursing assistant (CNA) interviews, and licensed nursing staff reports. Trained research staff collected all data. Periodic audits were conducted to assure that data collection procedures and scoring remained consistent across staff members. Proxy data were collected from nurses and CNAs who knew the resident well through direct caregiving or supervision of the care planning process. All measures for each resident were collected within one day before or after the resident interview and whenever possible, on the same day.

Criteria for study measures were that they had documented reliability and validity for older adults living in nursing homes as well as clinical utility, that is, they were brief and relatively easy-to-use. The rationale for this choice was twofold. First, several process and outcome measures were required to test the hypotheses for the parent grant; thus where possible, we chose briefer tools to measure individual variables to minimize respondent burden. Second, we wanted to use “ecologically valid” tools, that is, measures that could easily be used by clinicians during everyday care (37). In this way, the intervention and measures could be used in future effectiveness studies and quality improvement initiatives.

Demographic Variables

Information about the participants’s age, race, educational level and other demographic data were abstracted from the medical chart. The source document for most of this information was the admission MDS, a federally-mandated assessment and care planning tool (35). In a few instances, data in the MDS were incomplete and research staff examined other parts of the medical record (e.g., social work assessment) to locate the data.

Cognitive Status

Cognitive functioning was measured using the Cognitive Performance Scale (CPS), a 5-item instrument derived from the most recent MDS (38). Items measure short and long-term memory, cognitive skills for decision-making, communication, and independence in eating. Following standard decision rules, raters assign a summary score of 0 (intact) to 6 (very severe impairment). Morris et al. reported that scores on the CPS were significantly associated with the Mini-Mental State Examination (MMSE) scores, nursing assessments, and neurological diagnoses of Alzheimer’s disease and other dementias (38). Overall inter-rater reliability of all items is acceptable at 0.85. The CPS score was calculated using the participants’ most recent MDS from the medical record and was used to describe the sample. Ability to provide reliable self-report was determined using other methods (see below).

Self-Reported Pain Variables

1. Pain Intensity

Self-reported pain intensity was measured using the Iowa Pain Thermometer (IPT). The IPT uses a graphic representation of a thermometer in which the base is white and becomes increasingly red as one moves up. The base is anchored with the words “no pain,” and the top of the thermometer is anchored with “the most intense pain imaginable.”

Thirteen evenly spaced circles, corresponding with numeric values from “0” to “12,” are placed along the thermometer with verbal descriptors. Several studies have shown that the IPT is reliable, valid and generally preferred over other pain intensity tools (39–42). Herr et al. (39) compared five pain intensity tools in a sample of 36 older adults with chronic pain (22% with cognitive impairment) and found that the IPT had the lowest failure rate and was sensitive to change as demonstrated by significantly lower pain ratings following therapeutic intra-articular joint injections. For the current study, participants were handed a copy of the IPT that filled an 8½ × 11 inch card and asked to place a finger on the place on the IPT that corresponded to their usual, least and worst pain during the past week and the level of their pain now. Research staff then marked the number of the circle that most closely matched the point on the IPT indicated by the participant.

Only data from participants who could provide verbal, reliable responses were included in this study. Participants who were completely nonverbal or unable to respond to the question, “Have you experienced aches, pains or discomfort in the past week?” (alternative phrases, e.g., “Do you hurt anywhere?” Do you have any sore spots?” were used, when necessary) and to rate their pain intensity were listed as “nonverbal” and excluded from the analysis. Participants in the current study also had to give reliable responses. Reliability was assessed in one of two ways. When the study began, participants were asked to describe the worst pain they had ever experienced and locate its intensity on the IPT scale. A response was considered valid if the participant reported experiencing their worst point and located it on the top third of the IPT (all reported having experienced severe pain as their worst). Some participants, however, struggled with assessment even though they otherwise seemed capable of responding to other questions in the interview. Through discussions with the data collection staff and among the investigators, an alternate approach was taken. In this approach, the research assistant who interviewed the participant reviewed the participant’ answers for usual, worst, least and current pain. If all responses were consistent (e.g., worst pain was greater than least pain), then the participant was assessed as providing reliable answers. If responses were not consistent, the research assistant then asked the participant for clarification (e.g., “How is your least pain higher than your current pain?”). If the participant changed the response to a logically consistent one, then the response was marked as reliable. If the participant still couldn’t correct the discrepancy, the response was considered unreliable and the participant’s data were not included in the analyses. Reports of least and current pain were used to assess reliability of self-reported pain responses, but only the reports for usual and worst pain were analyzed in this study.

2. Pain Locations

To measure the number of painful body areas, participants were asked to respond “yes” or “no” as to whether they had pain, discomfort, achiness, or soreness in each of ten body areas. The research assistant recorded participants’ response.

3. Pain Pattern

Participants were given a response card and asked to report which of the following statements best described their pain (or “achiness,” “soreness,” “discomfort”) in the past week: “I have pain most of the time (constantly);” “I have pain most of the time but it’s sometimes worse than at other times;” “I have pain that comes and goes – at times I don’t have any pain;” “I have not had any pain in the past week.” Data collectors documented the participants’ response.

Painful Conditions

Data about painful conditions were collected from the medical record using the MDS and the resident’s medical problem list. Two doctorally-prepared gerontological nurse researchers, a gerontological nurse practitioner, and a geriatrician, each with expertise in pain management, reviewed participants’ diagnoses and judged whether each condition was consistently (i.e., in at least 75% of persons with the disorder) associated with pain. At least three of the four experts had to agree for a diagnosis to be accepted as painful. A final Painful Diagnoses score was calculated by adding the participant’s diagnoses judged as painful by the expert group. Diagnoses that appeared both in the MDS and problem list were counted only once.

Behavioral Indicators of Pain

Certified Nursing Assistant Observation of Behavioral Indicators of Pain. To identify pain-related behaviors, the CNA caring for the participant completed the Checklist of Nonverbal Indicators (CNPI). The CNPI assesses for 6 categories of pain-related behaviors: non-verbal vocalizations (e.g., moans, sighs), facial expressions (e.g., grimacing), bracing, restlessness, rubbing, and verbal complaints (e.g., “ouch,” “that hurts”). Although the CNPI focuses on a limited number of pain-related behaviors, these behaviors are likely most specific for pain (10).

CNAs were asked to observe the participant at rest and also during movement or transfers that occurred during morning care and report each behavior observed. To minimize missing data and to assure that the CNA completing the tool understood the measure, research assistants interviewed the CNA after the participants’ morning care and documented the behaviors reported by the CNA. The number of checks for each condition (i.e., at rest and during movement) is summed, thereby yielding a total of “0” to “12” observed behaviors. The CNPI has face validity (11, 43) and preliminary psychometric testing supports concurrent and construct validity as well as inter-rater reliability (11, 43). Although there are no rules or standards for interpreting the final score (11, 43), there is some empirical evidence that the number of pain behaviors observed is positively correlated with pain intensity (44). Prior to beginning data collection at every facility, CNAs viewed a brief instructional video about the CNPI and how to use it. Only CNAs who were familiar with the participant (i.e., had cared for the participant regularly in the past month) were asked to complete the CNPI. If on the data collection day the CNA assigned to the participant was unfamiliar with the participant, the interview was completed during another shift, or the research staff attempted to identify another CNA who was familiar with the participant or had cared for him/her in the previous 1 – 2 days.

Known Correlates of Pain

1. Depression

The Cornell Scale for Depression in Dementia (CSDD) was used to measure depression. The CSDD is a 19-item scale that includes information from semi-structured interviews with participants and their caregivers. Items are grouped under categories, “mood-related signs,” “behavioral disturbance,” “physical signs,” “cyclic functions,” and “ideational disturbance,” and are rated 0 = absent, 1 = mild or intermittent, or 2 = severe and then totaled to obtain a score. Scores above 10 indicate probable major depression (45). Although originally developed for persons with dementia, the CSDD has acceptable reliability and validity in individuals who are cognitively intact as well (45–49). Nurses overseeing the participant’s plan of care were interviewed as the caregiver. Data from interviews with each participant were not included because of difficulties interpreting participant’ behaviors given the research staff’s lack of prior interaction with, and knowledge of, the participant.

2. Agitation

Agitation was measured using the Pittsburgh Agitation Scale (PAS). The PAS is an observer-rating of four groups of behaviors: aberrant vocalization, motor agitation, aggressiveness, and resistance to care. Respondents rate the highest intensity score for each behavior group observed during the past week. The four items are scored on a 0 – 4 scale, with “0” indicating the behavior was not present and “4” indicating the most intense demonstration of the behavior. Overall scores range from 0--16, with higher scores denoting high levels of agitation. The PAS has acceptable reliability and validity (50) and is significantly associated with behavioral and surrogate pain measures in nursing home residents with moderate to severe cognitive impairment (26). For the current study, the PAS was completed by the nurse responsible for overseeing the participant’s overall plan of care.

Proxy Pain Reports

1. CNA Assessment of Residents’ Pain

Proxy pain scores were obtained from the CNAs using the IPT. After completing the CNPI, CNAs were asked to use the IPT to rate the residents’ usual and worst pain based on behaviors they observed. As with the self-report measure, scores ranged from 0 = “no pain” to 12 = “the most intense pain imaginable.” At the beginning of the study, CNAs reported on participant’ usual pain only. After recruiting the first four sites, the investigators added worst pain to the CNA-IPT measure. Because of this change, analyses involving the CNA-IPT involve a smaller sample size.

2. MDS Pain Variables

For the current study, we collected the MDS pain frequency item (J2a) from residents’ medical record. This item asks the nurse to think back on the previous 7 day period and assess the participant’s pain during this time as 0 = no pain, 1 = pain less than daily, or 2 = pain daily. We also collected the MDS pain intensity score (item J2b), which is recorded as missing = no pain (as indicated in item J2a), 1 = mild pain, 2 = moderate pain, and 3 = times when pain is horrible or excruciating. In addition to these two variables, we calculated an MDS Summary score in which the MDS pain frequency was multiplied by the MDS pain intensity, yielding a range of scores from 0 = no pain to 6 = daily pain that was at times horrible or excruciating. The MDS is completed by the nurse overseeing the residents’ care or by the facility nurse whose responsibility is to complete and submit MDS assessments for residents (i.e., MDS Coordinator). A registered nurse is required to complete the MDS, and instructions for completing the tool emphasize the need to communicate with and observe residents, review the medical record, and communicate with the interdisciplinary team to ensure accuracy in assessing residents (35). The MDS is completed for every resident on admission, with significant clinical changes, and at least quarterly.

Statistical Analyses

Descriptive statistics were calculated for all variables and presented as mean ± standard deviation (SD) for numerical variables and as percentages for categorical variables. The association of demographic and pain score variables with self-reported pain scores (self-reported usual and worst pain scores) was assessed using the Spearman correlation coefficient (ρ). We found no association of the nursing home site with either the usual or worst self-reported pain outcome (ICC = 0.00) and, therefore, site was not included in any of the modeling.

Multivariate models to predict the self-reported usual pain score and, separately, self-reported worst pain score were created using backward elimination of variables and, in a separate modeling exercise, forward selection of variables in least-squares linear regression. The forward and backward methods are commonly used due to their simplicity and are appropriate for exploratory analyses (51). The goal of the modeling, in our context, was to determine the potential of a set of independent variables to predict a specified outcome. Although the literature supports the inclusion of each independent variable in predicting pain self-report, we did not have a priori hypotheses about which combinations of independent variables would explain the most variance in the dependent variables; thus, our analytic approach was exploratory. Exclusion of variables in the backward or forward process does not mean that they are not causally related to the outcome. The variables that do remain in the final model may be useful for prediction. The P-values for variables finally retained after the forward and backward procedures, tend to be estimated as too small, due to the larger pool of variables from which the variables are selected at each step of the selection procedure.

All 13 variables considered in the correlation analysis were considered in the variable selection procedures and P<0.05 was used as the criterion for retention in the backward elimination and for inclusion in the forward selection procedure. All calculations were performed in the statistical language R, version 2.8.0 (Vienna, Austria). P < 0.05 was used to designate statistical significance.

Results

The final sample comprised 326 self-reporting participants with a mean age of 83.2 years. Sixty-nine percent were women, and 94% were non-Hispanic Whites. The most commonly reported sites of pain were legs and feet (82%) and back (61%). Forty-six percent had constant pain during the previous week, and the mean reported usual pain was at a moderate level (mean: 5.3, SD=2.3 on a 0 – 12 scale). Demographic and pain-related information, as well as means and standard deviations (SD) of study variables are presented in Table 1.

Table 1.

Sample Demographic and Pain Characteristics

n = 320–326, unless specified otherwise.

	Mean±SD or %
Age (yrs)	83.2±8.0
Female	69%
Education
<12 yrs	16%
High school graduate	40%
Some college, trade school	29%
Bachelor’s degree	13%
Graduate degree	3%
White, nonHispanic	94%
MDS Cognitive Performance Scale (possible range: 0–6)	2.1±1.3
MDS pain intensity score (possible range: 0–2)	1.3±1.0
MDS pain summary score (possible range: 0–6)	2.3±1.6
Cornell depression score (possible range: 0–38)	3.2±3.3
Pittsburgh Agitation Scor(possible range: 0–16)e	1.5±2.4
CNA usual pain in past week a (possible range: 0–12)	3.4±2.6
CNA worst pain in past week a (possible range: 0–12)	4.8±3.0
(CNA CNPI Total score possible range: 0–12)	3.8±2.9
Number of painful diagnoses (actual range: 0–7)	2.0±1.5
Pain Locations	%
n=325–326
Head	35%
Neck	48%
Back	61%
Shoulder	54%
Arms and hands	60%
Buttocks and hips	60%
Abdomen	29%
Legs and feet	82%
Chest	23%
All over	37%
Pain Pattern	%
Constant	6%
Constant plus periods of worsening	40%
Intermittent	53%
No pain	2%
Self-reported pain (0–12 scale)
n=313–326
Least pain	1.9±2.2
Worst pain	7.6±2.2
Usual pain	5.3±2.3
Current pain	3.4±3.0

^an = 256 for usual pain and 259 for worst pain.

Correlation Analysis

The results of the correlation analysis are presented in Figure 1. All MDS pain variables (frequency, intensity, summary score), and CNA IPT usual and worst pain scores were significantly associated with each of the two self-reported scores with correlations ranging from ρ = 0.13 to 0.26. The CSDD was significantly correlated only with worst self-report score (ρ =0.17). Spearman correlations for the remaining (non-significant) pain indicator predictors ranged between −0.08 and 0.10, suggesting only weak association or no association. For the demographic variables, only education level was significantly associated with self-reported usual pain (ρ = −0.17, P = 0.03).

Multivariate Model for Self-Reported Usual Pain

Both the backward and forward variable selection procedures yielded the same multivariate model for usual pain. The model is presented in Table 2. The final model includes the CNA IPT usual pain score, PAS, CSDD, and education level (R² = 0.14). An unexpected finding of the multivariate analysis was the significant negative association of the PAS summary score with self-reported usual pain score (Table 2.) However, this negative association was negligible in the univariate correlation analysis (Figure 2).

Table 2.

Multivariate Models for Self-Report Usual and Worst Pain

	Self-Report Usual Pain	Self-Report Worst Pain	Self-Report Worst Pain
	(n= 233)	(n = 244)a	(n = 244) b
Variable	coef (95% CI)±SE P	coef (95% CI)±SE P	coef (95% CI)±SE P
MDS Pain frequency			0.52 (0.15, 0.89)±0.19 0.006
MDS Pain intensity		0.30 (0.05, 0.54)±0.13 0.02
CSDD	0.14 (0.06, 0.23)±0.04 0.001	0.07 (0.00,0.13)±0.03 0.045
PAS summary score	−0.16 (−0.31, −0.01)±0.08 0.03
CNA IPT usual pain	0.18 (0.07, 0.28)±0.05 0.001
CNA IPT worst pain		0.24 (0.13, 0.36)±0.06 <0.001	0.25 (0.13, 0.37)±0.06 <0.001
CNPI summary score		−0.13 (−0.25, −0.01)±0.06 0.04	−0.13 (−0.25,−0.01)±0.06 0.04
Age (per 10 years)		−0.37 (−0.69, −0.05)±0.16 0.02	−0.38 (−0.70, −0.06)±0.16 0.02
Education c	−0.39 (−0.66, −0.13)±0.14 0.004
Intercept	5.40 (4.58, 6.22)±0.42	9.48 (6.76, 12.19)±1.38	9.67 (6.96, 12.37)±1.37
Model summary	R2 = 0.14, SEE= 2.13 (P<0.001)	R2 = 0.14, SEE=2.05 (P<0.001)	R2 = 0.13, SEE=2.06 (P<0.001)

Coef = linear regression coefficient; SEE = standard error of the estimate.

^aBackward elimination method utilized.

^bForward elimination method utilized.

^cEducation as a ranked predictor (ranging from <12 yrs =1 to graduate degree = 5, see Table 1 for remaining categories).

Figure 2.

Spearman correlation of predictors with usual and worst self-report pain. Corresponding sample size and P-values are shown at the bottom of the figure. MDS = Minimum Data Set; CSDD = Cornell Scale for Depression in Dementia; PAS = Pittsburgh Agitation Scale; CNA IPT usual pain = Certified Nursing Assistant proxy usual pain score; CNA IPT worst pain = Certified Nursing Assistant proxy worst pain score; CNPI Sum Score = Checklist of NonVerbal Pain Indicators total number of observed behaviors; CPS = Cognitive Performance Scale; #PD = number of painful diagnoses.

Multivariate Model for Self-Reported Worst Pain

Slightly different multivariate models were selected for self-report worst pain by the backward elimination and forward selection procedures (Table 2). The model selected by backward elimination included MDS pain frequency, CSDD, CNA IPT worst pain, CNPI summary score and age (R² = 0.14). The model selected by forward selection included MDS pain intensity, CNA IPT worst pain, CNPI summary score and age (R² = 0.13). The two models would be similar in their precision of prediction, given their similarity in values of R² and the standard error of the estimate (SEE). The SEE represents the variability of observations around their regression-predicted value.

Discussion

Since no single observer-recorded pain tool adequately predicts pain intensity, the aim of this study was to explore the contribution of several observer-recorded variables to predict self-reported pain in nursing home residents. Our results suggest that multiple indicators of pain, including agitation, depressive symptoms, and number of painful diagnoses, do not perform significantly better than a single measure—nursing assistant proxy report. These findings need to be understood in the context of existing models of pain in persons with dementia and the limitations inherent in any one study to address the complexities of measuring pain in population.

We examined usual and worst pain separately because persistent pain, which is common among nursing home residents, waxes and wanes. It is possible that observer-recorded pain variables, either singly or combined, could be associated inconsistently with usual and worst pain. For example, nurse reports may align more closely with usual pain because the nurse may not be present when residents transfer or ambulate, thereby missing many episodes of residents’ worst pain. Indeed, the models for usual and worst pain included different variables, although no model for either type of pain explained more than 14% of the variance.

The final model for self-reported usual pain used four variables

CNA IPT usual pain score, PAS, CSDD, and Education. The association of pain intensity and PAS agitation in the multivariate model was negative. This finding was unexpected, in that agitation often is considered a behavioral pain indicator in people with dementia (1, 5, 52) and other studies have demonstrated significant positive correlations between the two variables (26–28, 53). The negative association of the PAS score with self-reported usual pain may be due to chance. However, it may also suggest that cognitively intact individuals respond differently to pain than those who are severely cognitively impaired. Agitation and restlessness, for example, may exacerbate musculoskeletal pain, which is the most common source of pain in frail elders (31, 54). A person who is aware that movement increases pain (i.e., cognitively intact) will tend to stay relatively immobile to minimize discomfort. On the other hand, the person who is severely cognitively impaired may not be able to make the logical connection between movement and pain, and instead become restless and agitated. Similarly, the cognitively intact person is not likely to engage in other socially unacceptable behaviors – resistance to care and verbal aggressiveness – associated with agitation.

These unexpected results could cast doubt on using self-report as a criterion measure against which to evaluate the accuracy of observer-recorded variable. Other authors have identified only modest correlations between self-reported pain and direct observations of pain behavior (55), and self- and proxy pain reports (18, 19). We chose to use self-report as the dependent variable because it is generally considered the gold standard for assessing pain, and as such, is one validation paradigm for developing a pain measure in persons with advanced dementia who are nonverbal (34). However, this approach can only provide preliminary validation of the measure. Our findings support the use of additional, alternative analytic approaches using different criterion measures.

Using exploratory backward and forward variable selection procedures in regression to predict self-reported worst pain, we arrived at two very similar multivariate models. Both models included CNA IPT worst pain (the strongest predictor in the correlation analysis) the CNPI summary score, and age. The first model also included MDS pain intensity and CSDD and the second model, MDS pain frequency. The predictive performance of the two models was very similar (R² = 0.14 vs. R² = 0.13). The negative value for the CNPI summary score coefficient was unexpected, since we would have predicted a positive association between the number of observed pain behaviors and self-reported pain. However, similar explanations may be proposed as for the negative associations between agitation (PAS) and self-reported usual pain. First, the negative associations found in the regression analyses may be spurious. This is likely, given the positive correlations between the CNPI summary score and self-reported pain that were found in the univariate analyses. Second, there may be differences in behavioral responses to pain between persons who are cognitively impaired and those who are cognitively intact as suggested above. Finally, the results may be reflective of inherent problems with the CNPI. Since the study was initiated several years ago, investigators have developed and tested other pain behavioral observation tools; some of these tools have been found to have stronger psychometric properties and are more widely used than the CNPI (12, 56, 57). Our own research has revealed other psychometric limitations of the CNPI as well as the PAINAD (58).

The self-reported pain for individual residents at a single time may not be well predicted by our models, as indicated by the SEE values of approximately 2 from the regression models in Table 2. (The SEE is approximately the standard deviation of residuals around the regression-predicted pain score.) Using two standard deviations as a plausible range of variation, the SEE of 2 implies that residents may commonly vary by four points (on a 10-point scale) around the predicted value from the models. Nevertheless, the average pain for a sufficiently large group of residents may be accurately predicted from the models. Such averages may be useful in intervention or treatment trials where the primary outcome is the change in a mean over time or a comparison of means or mean changes between two groups. Also, the average pain of an individual resident over a period may be well predicted if there are a sufficient number of independent observations of that resident during the period.

Our multivariate models may show a better prediction (R²) of self-reported pain than would be encountered with a new sample. The backward elimination method, starting with more than ten candidate variables for entry into a regression model, allows more than a 0.05 chance for a “false positive” entry of a variable into the model. The false positive would be a predictor variable that is correlated with self-reported pain in our sample, but whose population correlation is zero or close to it. Nevertheless, the quite small P-values for the regression models (Table 2) do suggest that some of the associations noted have a good chance of being true associations in the population of nursing home residents.

The statistically significant correlations between CNAs’ estimates of residents’ pain and residents’ self-report support the findings of other investigators (59, 60). Most of the direct caregiving in nursing homes is done by CNAs and this intense interaction may enhance their ability to observe and estimate the pain experienced by residents. However, the correlations were modest and suggest that even the “best” predictor for self-reported pain is inadequate. More intensive CNA training in observing pain and making a proxy judgment about pain may increase the concordance between their estimates of residents’ pain (18, 19).

This study was exploratory and its limitations must be acknowledged. First, we attempted to include data only from participants able to reliably self-report; however, there is no widely tested and reliable method for determining the reliability of these responses and our chosen methods may have incorrectly categorized some participants. Second, our sample was largely comprised of white, non-Hispanic participants. Although this profile matches that of the population in many nursing homes, cultural and gender differences in pain reporting may have been missed. Also, some of the pain variables showed statistically significant differences among facilities, most likely due to true differences in the average severity of pain of participants at the different facilities. However, it is also possible that either 1) different record-keeping and research response practices at facilities obscured true associations among pain and other variables or 2) that some of the associations that we noted among variables are due to facility effects and unmeasured facility-related variables. There is also a limitation inherent in the variable selection procedures utilized in the multivariate analysis. While it is valid to use selection procedures, these procedures maximize prediction and may sometimes lead to overfitting of the model. Finally, the concept behind each of the predictor variables, such as agitation or depression, may be closely related to pain, but the conversion of these concepts into numerical scores is not perfect. Thus, it is possible that improvements in measuring pain-related concepts will yield a better predictive model for self-reported pain.

Despite its drawbacks, the study has a number of strengths. It was carried out in 24 facilities that represent a diversity of bed size and quality of care. The recruitment screened carefully for residents that did, indeed, have pain, and the average pain scores do show that pain is a serious issue with these residents. The data collection was centrally directed by a staff that had little turnover, lessening the impact of person-to-person variation in pain assessment. The assessment of reliability of self-reported pain, while not perfect, was used to substantially limit the number of meaningless or erroneous self-reports of pain which were included in the analysis. We used a number of standardized and well-established pain and other scales in the search for meaningful correlates of self-reported pain. And the sample size of 233 residents included in most of the analyses was sufficient to detect relatively weak correlation coefficients—around 0.2 or larger at 80% power.

Findings from this study underscore the challenges that caregivers and researchers face in measuring and assessing pain in persons who are unable to self-report. Estimates of pain from a single source do not adequately predict self-reported pain, although in this study, multivariable models only slightly improved predictive accuracy. Future efforts must focus on identifying the best variables or combination of variables to estimate pain when the person is unable to self-report; these composite pain measures must be evaluated against several criterion measures as self-reported pain may be unreliable or unavailable (18, 34). Despite the complexities of developing and evaluating multidimensional pain measures for persons who are unable to report pain, novel analytic approaches must be used to address this challenge. Otherwise, the under-assessment and inadequate treatment of pain in this vulnerable population will continue, causing untold and unnecessary suffering.